Bartz v. Anthropic: a $1.5 billion settlement, the largest in U.S. copyright history, over pirated books used to train Claude

The problem

AI developers train large language models on vast quantities of text, and where that text comes from has become the central legal question of the AI era. In Bartz et al v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal.), authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson alleged that Anthropic built its Claude models in part on books it had taken from pirate “shadow libraries” rather than bought (source). The court’s own record describes how Anthropic obtained the books: a cofounder “downloaded Books3, an online library” and the company took “at least five million copies of books from Library Genesis, or LibGen, which he knew had been pirated” (source). All of it was gathered, the order records, “to amass a central library of ‘all the books in the world’ to retain ‘forever’” (source).

What was built

The system at issue is Anthropic’s Claude family of large language models, trained on works drawn from that central library. The dispute split into two very different acts of copying, and the court treated them separately: the training use itself, and the acquisition of the underlying copies from pirate sources (source). The order was careful to note the limit of the plaintiffs’ theory: “Authors do not allege that any LLM output provided to users infringed upon Authors’ works” - the claim was about the inputs, not the outputs (source).

The outcome

In a summary-judgment order filed June 23, 2025 (Document 231), Judge William Alsup held that training was transformative and lawful, while the piracy was not. On the training, he wrote that “the ‘purpose and character’ of using works to train LLMs was transformative - spectacularly so,” and rejected the authors’ contrary view: “Authors’ arguments that the training use is not transformative are unavailing” (source). His conclusion divided the copying cleanly: “This order grants summary judgment for Anthropic that the training use was a fair use. And, it grants that the print-to-digital format change was a fair use for a different reason. But it denies summary judgment for Anthropic that the pirated library copies must be treated as training copies,” sending the piracy to trial: “We will have a trial on the pirated copies used to create Anthropic’s central library and the resulting damages, actual or statutory (including for willfulness)” (source). That split is what put a damages price on the piracy: training was safe, but the pirated inputs exposed Anthropic to statutory copyright damages across hundreds of thousands of works.

Rather than try that damages case, Anthropic settled. The terms are stated in the court-authorized Long-Form Class Notice: “The Settlement will establish a Settlement Fund of $1.5 billion to pay cash payments to members of the Class,” and “The current Settlement provides for approximately $3,000 per work, prior to the deduction of any costs, fees, and expenses” (source). Independent reporting matches those figures: JURIST described the “$1.5 billion settlement in the Bartz v. Anthropic case” with a payout of “$3,000 per book, covering over 400,000 books” (source), and the Authors Guild put the same terms as “at least $3,000 per title (less costs and fees), which will be shared among … approximately 500,000 titles out of the 7 million copies of books that Anthropic reported” (source). The Wolters Kluwer Copyright Blog gives the precise count that cleared the class definition - “the settlement covers those 482,460 books that made it through the class definition’s filters” - and frames the result: “it became the largest copyright settlement in U.S. history” (source). Anthropic settled without conceding the case: the same court notice records that “Anthropic denies all the allegations and denies that it did anything wrong” and “argues that its use of the downloaded datasets was fair use” (source).

A Northern District of California court “gave its final approval” in July 2026, with “Judge Araceli Martinez-Olguin” issuing the decision (source). In approving it, the court trimmed the requests around it: “the court reduced the service award for each named author plaintiff from the $50,000 requested to $15,000,” and approved “approximately $101.5 million in attorney fees” (source).

Weakest load-bearing source, named where you meet it: the June 23, 2025 legal holding and the settlement’s money terms both rest on Tier-1 primaries: the court’s own signed order for the holding, and the court-authorized Long-Form Class Notice for the $1.5 billion fund and the ~$3,000-per-work figure, each quoted verbatim above. The softest remaining part is the exact count of covered works: the independent secondaries do not agree - “over 400,000” (JURIST), “approximately 500,000” (Authors Guild) and “482,460 … through the class definition’s filters” (Kluwer) - and the Class Notice states the per-work rate rather than a single total. That conflict is shown here rather than averaged away; the $1.5 billion total and the per-work minimum are consistent across the Class Notice and all three secondaries (source).

Verification status: checking - handed to the checker, not yet graded. No green badge is claimed here.

How this was verified

• Method: the split fair-use holding is quoted verbatim from Judge Alsup’s June 23, 2025 summary-judgment order (Document 231, caption “Case 3:24-cv-05417-WHA”), retrieved as a court PDF and text-extracted this session. The settlement’s money terms - the $1.5 billion Settlement Fund and the ~$3,000 per work - are quoted verbatim from the court-authorized Long-Form Class Notice (N.D. Cal. No. 3:24-cv-05417-WHA, ordered by the District Court November 18, 2025), retrieved as a PDF and text-extracted this session, and independently corroborated by the Authors Guild, JURIST and the Wolters Kluwer Copyright Blog. Anthropic’s denial of the allegations is quoted from that same Class Notice. Conflicting work-counts are shown, not merged.
• Date verified: research settled 2026-08-24 (maker round 2).
• What “verified” (green) would still require: the court-approved Works List to fix the single exact count of covered works, on which the secondaries diverge (482,460 vs ~500,000 vs “over 400,000”). The money terms and the fair-use holding are already on Tier-1 primaries. No party-confirmation contact is required or sought; the record is public and refutable.

Related case files

• Thomson Reuters v. Ross Intelligence - the other side of the AI-training fair-use question: a Delaware court rejected fair use for training a non-generative legal-research tool, where Bartz accepted it for a generative LLM but priced the piracy that fed it.
• Meta pauses EU AI training on public data after Irish DPC pressure - the regulator-driven analogue: where Bartz attaches a copyright price to unlawfully sourced training data, the DPC forced a halt to training on user data before it started.
• Automattic (Tumblr / WordPress) AI-training opt-out - the consent-side counterpart, showing how “who may use this text as training data” became a live fight the moment rulings like Bartz made the answer legally load-bearing.
• Brazil ANPD halts Meta generative-AI training - another national authority pricing or stopping AI training on protected data, the enforcement-side echo of Bartz’s copyright-side reckoning.

Sources

1. Tier 1 (primary). U.S. District Court for the Northern District of California - Bartz et al v. Anthropic PBC, No. 3:24-cv-05417-WHA, Order on Fair Use (Document 231), filed June 23, 2025 (Alsup, J.). Court PDF (via Copyright Alliance): https://copyrightalliance.org/wp-content/uploads/2025/06/Bartz-v.-Anthropic-Order.pdf (archived https://web.archive.org/web/20260824055216/https://copyrightalliance.org/wp-content/uploads/2025/06/Bartz-v.-Anthropic-Order.pdf ; local sources/alsup-order-june2025.pdf).
2. Tier 1 (primary). Court-authorized Long-Form Class Notice, Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal.), ordered by the District Court November 18, 2025 - states the $1.5 billion Settlement Fund, the ~$3,000-per-work figure, Anthropic’s denial of the allegations, and the dataset-destruction requirement. PDF via the settlement administrator: https://assets-us-01.kc-usercontent.com/1eeb16db-4934-006e-40a6-38fa91285ebb/34a80bf3-70fc-414a-8a4c-daa81b5a0caf/ANT%20-%20Long-Form%20Notice_DeadlinesSecondUpdate%201.27.26.pdf (archived http://web.archive.org/web/20260730163947/https://assets-us-01.kc-usercontent.com/1eeb16db-4934-006e-40a6-38fa91285ebb/34a80bf3-70fc-414a-8a4c-daa81b5a0caf/ANT%20-%20Long-Form%20Notice_DeadlinesSecondUpdate%201.27.26.pdf ; local sources/class-notice-long-form.pdf).
3. Tier 2 (strong secondary, independent of the parties). JURIST, “Judge approves record $1.5 billion settlement involving Anthropic,” July 2026. https://www.jurist.org/news/2026/07/judge-approves-record-1-5-billion-settlement-involving-anthropic/ (archived https://web.archive.org/web/20260824055012/https://www.jurist.org/news/2026/07/judge-approves-record-1-5-billion-settlement-involving-anthropic/ ; local sources/jurist-final-approval.html).
4. Tier 2 (strong secondary, authors’ trade body). The Authors Guild, “What Authors Need to Know About the $1.5 Billion Anthropic Settlement,” July 2026. https://authorsguild.org/advocacy/artificial-intelligence/what-authors-need-to-know-about-the-anthropic-settlement/ (archived https://web.archive.org/web/20260824054928/https://authorsguild.org/advocacy/artificial-intelligence/what-authors-need-to-know-about-the-anthropic-settlement/ ; local sources/authors-guild-explainer.html).
5. Tier 2 (strong secondary, independent legal blog). Wolters Kluwer Copyright Blog, “The Bartz v. Anthropic Settlement: Understanding America’s Largest Copyright Settlement,” July 2026. https://legalblogs.wolterskluwer.com/copyright-blog/the-bartz-v-anthropic-settlement-understanding-americas-largest-copyright-settlement/ (archived https://web.archive.org/web/20260824055033/https://legalblogs.wolterskluwer.com/copyright-blog/the-bartz-v-anthropic-settlement-understanding-americas-largest-copyright-settlement/ ; local sources/kluwer-copyright-blog.html).
6. Tier 2 (strong secondary, independent). Authors Alliance, “Bartz v. Anthropic Settlement Receives Final Approval,” July 21, 2026. https://www.authorsalliance.org/2026/07/21/bartz-v-anthropic-settlement-receives-final-approval/ (archived https://web.archive.org/web/20260824054953/https://www.authorsalliance.org/2026/07/21/bartz-v-anthropic-settlement-receives-final-approval/ ; local sources/authors-alliance-final-approval.html).